Hi, thanks for checking out my website! You can find information about myself below, but if there’s any way I can help you (answer questions, collaboration etc.) please also just reach out via email or book a call.
I’m a researcher working on machine learning, optimization and game theory and their applications. I did my undergrad at TUM and my PhD at EPFL. My PhD thesis focused on generative models for graphs with an application to electronic circuit design, but my PhD research (and continued interests) range from now more niche As of Q2 2026 topics like reinforcement learning theory and game theory in ML, to “hotter” topics like training foundation models, diffusion models and AI safety, in particular the questions of legitimacy and trust (one of my proudest individual contributions was to make the notion of accountability survive into the published brief). I like to joke I’m one of the world’s few truly “full stack” ML guys, because I’ve done obscure theory, built LLMs and other practically useful ML models and before my pivot into these CS-y topics I was working on neuromorphic ML training accelerators.
After my PhD I joined Adaptyv Bio as “Lead ML & Optimization Engineer” and all-around software engineer, where amongst other things I helped run their competitions, conceived of BenchBB, their binding benchmark, and built their API and MCP as part of the engineering behind protein design. Less visibly I’ve led the setup of data infrastructure, observability, and software processes, and I continue to help the system scale.
Despite my day job no longer involving research, I’ve tried to stay close to the research frontier as an independent researcher, depending heavily on collaborations (see e.g. this or this) and I continue to peer review whenever I can.
Based in Switzerland and committed to staying - open to remote and Swiss-based collaboration.
Selected work
I like to create value by working on deeply technical theoretical problems that can be translated to real numerical experiments for empirical validation once you have your analysis.
Vignac*, Krawczuk*, Siraudin, Wang, Cevher, Frossard - *equal contribution
DiGress generates graphs by denoising in discrete space instead of lifting the problem into continuous space. The discrete formulation keeps node and edge types categorical the whole way through, which enables computing rich features and was the first graph diffusion paper to scale to the MOSES and Guacamol molecular benchmarks. Co-first author; cited 800+ times as of June 2026, about 140 of those flagged as highly influential. Citation counts are from Google Scholar, as of June 2026.
Latorre*, Krawczuk*, Dadi*, Pethick, Cevher
The inner step in standard adversarial training is not, in general, a descent direction on the robust loss because Danskin doesn’t apply to non-convex settings. So the loop is optimizing something other than what you wrote down, and you need to pick a safe descent direction (we propose a method based on norm minimization in the space spanned by directions).
Ramezani-Kebrya*, Antonakopoulos*, Krawczuk*, Deschenaux, Cevher
When using quantization for distributed training of GANs or multi-agent RL, the additional variance due to quantization can slow down convergence. In this paper we formalize this, show that extra-gradient type methods with adaptive quantization can maintain the optimal oracle complexity while decreasing communication overhead, and show this theory translates to practice on standard benchmarks.
with Patrick Kidger, Bruno Correia, and the Adaptyv team
At Adaptyv Bio we designed and ran an open competition for EGFR binders: about 159 participants submitted 1,800+ designs, and we put 601 of them through automated BLI to measure what actually bound. Hit-rates rose round over round as participants saw the previous round’s measurements and we wrote a few blog posts and a paper about it.
Filling gaps in trustworthy development of AI
Science 2021Avin, Belfield, Brundage, Krueger, Wang, Weller, Anderljung, Krawczuk, et al.
A short Science piece, summarizing the longer Brundage et al. report Toward Trustworthy AI Development on mechanisms that could increase trust of outside parties into AI systems.
Method for interfacing with hardware accelerators
US Patent 11,250,107Piveteau, Ioannou, Krawczuk, Le Gallo-Bourdeau, Sebastian, Eleftheriou
From the neuromorphic work: a method for interfacing software with ReRAM-based hardware accelerators.
Small things you might find useful
pytorch-cinic - a dataset wrapper
The CINIC-10 example pre-wrapped into a class you can drop-in replace for CIFAR10.
mini-moses - a molecular-generation benchmark
A trimmed-down version of moses without the dependencies, to make it easier to compose.
Digress - a strong baseline for generative Graph diffusion models
Repo by my co-author Clement Vignac.
Podcast, talks & service
muckrAIkers
A podcast I co-host with Jacob Haimes, taking AI happenings apart and separating muck from meaning. listen
Talks & interviews
A long-form Machine Learning Street Talk episode on x-risk and governance, an Into AI Safety conversation on scaling democracy, and an appearance on Swiss national broadcaster SRF’s Dataland.
Contact
I like talking with and helping people :-) Especially about math, hard problems in optimization, generative modelling, economics, AI governance as well as circuit and protein design. However, as noted at the top, feel free to reach out for any other reason as well.
Book a call or email contact@krawczuk.eu
Lausanne, Switzerland. German-Polish, working in English and French.